We propose a novel self-supervised learning method for medical image analysis. Great progress has been made in medical image analysis because of the development of supervised learning based on deep convolutional neural networks. However, annotating complex medical images usually requires expert knowledge, making it difficult for a wide range of real-world applications ($e.g.$, computer-aided diagnosis systems). Our self-supervised learning method introduces a cross-view loss and a cross-model loss to solve the insufficient available annotations in medical image analysis. Experimental results show that our method can achieve high detection performance for gastritis detection with only a small number of annotations.